Segment point clouds with different point numbers

[ZimengTan]

Hi,
How to inference point clouds with different point numbers for part segmentation or semantic segmentation?
The number of points in my dataset is around 15,000. I randomly sampled 2048 points at a time during training.

When testing, I want to get the predicted label for each point.
Should I feed the whole point cloud into the network at once (with batchsize=1)?
Will the difference in points between training and testing (2048 vs 15,000) damage performance?
Or should I sample 2048 points at a time until all points are predicted?

Thank you so much!